Role of large language models for etiological classification of brain stroke based on MRI brain reports: a feasibility study

Significantly impacting public health, ischemic stroke stands as one of the leading causes of disability and one of the primary causes of mortality in developed countries. According to the Global Burden of Disease, ischemic stroke is not only the second leading cause of death but also one of the major contributors to disability worldwide [1]. The diversity of ischemic stroke subtypes, with different etiologies and demanding distinct management strategies, requires an accurate etiological classification [2] without having found, to our best knowledge, literature regarding the presence and description of the possible etiology of stroke in radiological reports. Diverse classification systems, such as the TOAST (Trial of Org 10,172 in Acute Stroke Treatment) [3], CCS (Causative Classification of Stroke) [4] and ASCO (Atherosclerosis, Small vessel disease, Cardiac source, Other) [5], have been established to categorize these subtypes of stroke. Not all these classification systems are used in equal proportion: the TOAST classification, despite having the lower inter-rater reliability rate [6], is widely adopted in both clinical practice and research due to its simplicity. Both the CCS [7] and ASCO [8] classifications exhibit higher inter-rater reliability, but their utilization is less prevalent due to their complexity [9] (Fig. 1).

The use of an etiological stroke classification system through NLP tools applied to brain MRI reports can offer significant advantages in patient management, reducing diagnostic times, suggesting additional diagnostic tests focused on suspected etiology, and expediting the initiation of optimal treatment. It is also important to note that there are centers where brain MRIs in emergent contexts are not interpreted by experienced neuroradiologists, but by staff who may not be familiar with updated classification systems. This tool can provide substantial support by suggesting potential etiology with the aforementioned advantages, as well as improving interreader reproducibility.

At this point, from the radiologist's perspective, manual labeling of stroke etiologies searching for data from brain MRIs in patients' medical records can be tedious, time-consuming, prone to human error, and may result in low inter-rater agreement rates [10]. Artificial Intelligence (AI) based algorithms may help in this challenging and relevant task. Most of the AI solutions developed to date for stroke are focused on the analysis of radiological images (usually CT or MRI), offering variable accuracy rates not only in terms of stroke detection but also regarding stroke classification [11]. On the other hand, text-based AI approaches using natural language processing (NLP) technologies have opened new avenues for the automated extraction of valuable insights from radiological reports [12]. NLP, a branch of AI, can provide machines with the capability to understand human language (written or spoken). Several studies have previously evaluated the role of NLP in extracting and processing information from radiology and electronic health reports in several clinical scenarios with promising results [[13], [14], [15], [16]]. On the other hand, Large Language Models (LLMs) are sophisticated systems pre-trained on large amounts of text data to perform a wide variety of NLP tasks. There are a variety of LLMs available for use such as the BERT and RoBERTa models [17,18].

With the present study, our goal is to develop and validate an NLP algorithm using pre-trained language models able to automatically extract information from MRI brain reports in patients with suspected stroke and classify them, according to a modification of the TOAST classification, as lacunar, atherothrombotic, cardiogenic, or indeterminate to assist radiologists in real-time at their MRI reporting step. In other words, by harnessing the power of NLP, we aim to extract this hidden knowledge from the rich textual data found in radiological reports, ultimately enhancing our ability to accurately categorize stroke.

Comments (0)

No login
gif